Merge branch 'master' of master.kernel.org:/pub/scm/linux/kernel/git/davem/net-2.6
[linux-2.6] / kernel / cpu_acct.c
1 /*
2  * kernel/cpu_acct.c - CPU accounting cgroup subsystem
3  *
4  * Copyright (C) Google Inc, 2006
5  *
6  * Developed by Paul Menage (menage@google.com) and Balbir Singh
7  * (balbir@in.ibm.com)
8  *
9  */
10
11 /*
12  * Example cgroup subsystem for reporting total CPU usage of tasks in a
13  * cgroup, along with percentage load over a time interval
14  */
15
16 #include <linux/module.h>
17 #include <linux/cgroup.h>
18 #include <linux/fs.h>
19 #include <linux/rcupdate.h>
20
21 #include <asm/div64.h>
22
23 struct cpuacct {
24         struct cgroup_subsys_state css;
25         spinlock_t lock;
26         /* total time used by this class */
27         cputime64_t time;
28
29         /* time when next load calculation occurs */
30         u64 next_interval_check;
31
32         /* time used in current period */
33         cputime64_t current_interval_time;
34
35         /* time used in last period */
36         cputime64_t last_interval_time;
37 };
38
39 struct cgroup_subsys cpuacct_subsys;
40
41 static inline struct cpuacct *cgroup_ca(struct cgroup *cont)
42 {
43         return container_of(cgroup_subsys_state(cont, cpuacct_subsys_id),
44                             struct cpuacct, css);
45 }
46
47 static inline struct cpuacct *task_ca(struct task_struct *task)
48 {
49         return container_of(task_subsys_state(task, cpuacct_subsys_id),
50                             struct cpuacct, css);
51 }
52
53 #define INTERVAL (HZ * 10)
54
55 static inline u64 next_interval_boundary(u64 now)
56 {
57         /* calculate the next interval boundary beyond the
58          * current time */
59         do_div(now, INTERVAL);
60         return (now + 1) * INTERVAL;
61 }
62
63 static struct cgroup_subsys_state *cpuacct_create(
64         struct cgroup_subsys *ss, struct cgroup *cont)
65 {
66         struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
67
68         if (!ca)
69                 return ERR_PTR(-ENOMEM);
70         spin_lock_init(&ca->lock);
71         ca->next_interval_check = next_interval_boundary(get_jiffies_64());
72         return &ca->css;
73 }
74
75 static void cpuacct_destroy(struct cgroup_subsys *ss,
76                             struct cgroup *cont)
77 {
78         kfree(cgroup_ca(cont));
79 }
80
81 /* Lazily update the load calculation if necessary. Called with ca locked */
82 static void cpuusage_update(struct cpuacct *ca)
83 {
84         u64 now = get_jiffies_64();
85
86         /* If we're not due for an update, return */
87         if (ca->next_interval_check > now)
88                 return;
89
90         if (ca->next_interval_check <= (now - INTERVAL)) {
91                 /* If it's been more than an interval since the last
92                  * check, then catch up - the last interval must have
93                  * been zero load */
94                 ca->last_interval_time = 0;
95                 ca->next_interval_check = next_interval_boundary(now);
96         } else {
97                 /* If a steal takes the last interval time negative,
98                  * then we just ignore it */
99                 if ((s64)ca->current_interval_time > 0)
100                         ca->last_interval_time = ca->current_interval_time;
101                 else
102                         ca->last_interval_time = 0;
103                 ca->next_interval_check += INTERVAL;
104         }
105         ca->current_interval_time = 0;
106 }
107
108 static u64 cpuusage_read(struct cgroup *cont, struct cftype *cft)
109 {
110         struct cpuacct *ca = cgroup_ca(cont);
111         u64 time;
112
113         spin_lock_irq(&ca->lock);
114         cpuusage_update(ca);
115         time = cputime64_to_jiffies64(ca->time);
116         spin_unlock_irq(&ca->lock);
117
118         /* Convert 64-bit jiffies to seconds */
119         time *= 1000;
120         do_div(time, HZ);
121         return time;
122 }
123
124 static u64 load_read(struct cgroup *cont, struct cftype *cft)
125 {
126         struct cpuacct *ca = cgroup_ca(cont);
127         u64 time;
128
129         /* Find the time used in the previous interval */
130         spin_lock_irq(&ca->lock);
131         cpuusage_update(ca);
132         time = cputime64_to_jiffies64(ca->last_interval_time);
133         spin_unlock_irq(&ca->lock);
134
135         /* Convert time to a percentage, to give the load in the
136          * previous period */
137         time *= 100;
138         do_div(time, INTERVAL);
139
140         return time;
141 }
142
143 static struct cftype files[] = {
144         {
145                 .name = "usage",
146                 .read_uint = cpuusage_read,
147         },
148         {
149                 .name = "load",
150                 .read_uint = load_read,
151         }
152 };
153
154 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cont)
155 {
156         return cgroup_add_files(cont, ss, files, ARRAY_SIZE(files));
157 }
158
159 void cpuacct_charge(struct task_struct *task, cputime_t cputime)
160 {
161
162         struct cpuacct *ca;
163         unsigned long flags;
164
165         if (!cpuacct_subsys.active)
166                 return;
167         rcu_read_lock();
168         ca = task_ca(task);
169         if (ca) {
170                 spin_lock_irqsave(&ca->lock, flags);
171                 cpuusage_update(ca);
172                 ca->time = cputime64_add(ca->time, cputime);
173                 ca->current_interval_time =
174                         cputime64_add(ca->current_interval_time, cputime);
175                 spin_unlock_irqrestore(&ca->lock, flags);
176         }
177         rcu_read_unlock();
178 }
179
180 struct cgroup_subsys cpuacct_subsys = {
181         .name = "cpuacct",
182         .create = cpuacct_create,
183         .destroy = cpuacct_destroy,
184         .populate = cpuacct_populate,
185         .subsys_id = cpuacct_subsys_id,
186 };